Speakers - 2026

Stroke is a leading cause of disability worldwide and a major cause of long-term disability. Functional recovery and length of hospital stay (LOS) in inpatient rehabilitation are important indicators of patient outcomes, rehabilitation efficiency, and resource use. However, existing prediction studies based on traditional models have limited generalizability because of sample heterogeneity and differences in statistical methods. In addition, most studies rarely use data-driven machine learning (ML) approaches on large clinical datasets. This study aimed to address these gaps by developing prediction models using ML techniques on stroke rehabilitation outcomes from a large stroke registry in a local public hospital.

Methods:

A retrospective analysis was conducted on 5,133 patients in the stroke registry admitted to the Occupational Therapy Department of Tai Po Hospital between 2008 and 2021. The primary outcomes were Modified Barthel Index (MBI) gain, post-stroke disability level at discharge, and LOS after rehabilitation. The secondary outcomes were discharge disposition and inpatient mortality during rehabilitation. Predictors covered six domains: demographics, socioeconomic and medical information, and physical, cognitive, and functional performance. Multivariate linear regression was used to examine MBI gain and LOS, while multinomial logistic regression was used to investigate disability at discharge, discharge disposition, and inpatient mortality. Classification and regression tree (CART) analysis was used to identify decision rules and threshold values for the primary outcomes.

Results:

The mean MBI gain was 10.12 ± 19.61 and the mean LOS was 25.15 ± 18.11. Linear regression for functional gain showed excellent fit but appeared overfit (Adjusted R2 = 0.9996), whereas the LOS model showed moderate fit (Adjusted R2 = 0.250). Key predictors of MBI gain included admission and discharge MBI scores, receptive impairment at discharge, and first-ever stroke. The disability model showed discharge Functional Independence Measure (FIM) mobility score (B = −0.213 to −0.178, p < .01) and Modified Functional Ambulatory Category (MFAC) (B = −48.832 to −39.859, p < .001) as the strongest predictors. LOS was associated with balance (B = −0.150, p = .005), mobility level (B = 2.705 to 3.305, p < .05), and functional status (B = −4.166, p = .002). Predictors of discharge disposition to residential institution were multifactorial and included socioeconomic (B = 0.554, p = .007) and biomedical (B = 1.091 to 1.098, p < .01) factors. The CART model for functional gain showed satisfactory performance (test R2 = 0.811, MAE = 4.633, MSE = 36.548). The LOS CART model showed modest performance (test R2 = 0.255, MAE = 10.41, MSE = 227.928), suggesting that LOS was less amenable to regression tree prediction and likely influenced by broader clinical and non-clinical factors.

Conclusion:

The CART models provided interpretable clinical thresholds and simple decision rules for clinical decision making. They complemented traditional regression analysis and supported more structured data-driven inpatient rehabilitation planning. The findings may help frontline healthcare professionals with goal setting, discharge planning, and resource allocation for extended care in patients with stroke.

What will the audience take away from presentation?

• This study emphasizes the adoption of more structured, data-driven approach in formulation of treatment plans for patients with realistic goals and outcomes
• The CART-derived decision threshold also offers frontline staff with simple rules and logic in more justifiable clinical decision-making
• With the utilization of ML algorithm on routinely collected data in the stroke registry, complex statistical results could be converted into simple decision rules for daily evidence-based application

How will this help the audience in their job? Is this research that other faculty could use to expand their research or teaching? Does this provide a practical solution to a problem that could simplify or make a designer’s job more efficient? Will it improve the accuracy of a design, or provide new information to assist in a design problem? List all other benefits.

• This study served as a foundation study for evidence-based stroke rehabilitation outcome evaluation with application of data-mining techniques in Hong Kong. Ultimately, the clinical threshold derived from the
• CART models enable healthcare professionals, especially occupational therapists to deliver targeted and personalized care in stroke rehabilitation for patients with justifiable rationale and allocation of resources.